Re-express the calibrated NVFP4 components in NVIDIA Model Optimizer's unified Hugging Face layout: packed E2M1 values under `.weight`, the FP32 weight scale as `.weight_scale_2`, the static activation scale as `.input_scale` (fp32(amax)/2688), and a `quantization_config` (`quant_method: modelopt`, `quant_algo: NVFP4`) in each quantized component's config.json. Values and scales are unchanged bit for bit; the packed layout's manifests are removed. Produced by UniServe's scripts/convert_h3_modelopt_checkpoint.py (commit c002a175).
30 days ago
Document self-describing loading and generation validation
